Rework inception block tests

examples/inception_v3/tests/inception_a.py

@@ -1,62 +0,0 @@
-import sys
-
-import torch
-from torchvision.models.inception import InceptionA
-
-sys.path.append("../../../tools")
-from utils import print_cpp_vector
-
-torch.manual_seed(0)
-
-@torch.no_grad()
-def init_weights(m):
-    if isinstance(m, torch.nn.Conv2d):
-        torch.nn.init.uniform_(m.weight)
-    elif isinstance(m, torch.nn.BatchNorm2d):
-        torch.nn.init.uniform_(m.weight)
-        torch.nn.init.uniform_(m.bias)
-
-with torch.no_grad():
-    inception_a = InceptionA(3, 6)
-    inception_a.apply(init_weights)
-
-    # branch1x1
-    print_cpp_vector(torch.flatten(inception_a.branch1x1.conv.weight), "branch1x1_conv_weights")
-    print_cpp_vector(torch.flatten(inception_a.branch1x1.bn.weight), "branch1x1_bn_weights")
-    print_cpp_vector(torch.flatten(inception_a.branch1x1.bn.bias), "branch1x1_bn_bias")
-
-    # branch5x5
-    print_cpp_vector(torch.flatten(inception_a.branch5x5_1.conv.weight), "branch5x5_1_conv_weights")
-    print_cpp_vector(torch.flatten(inception_a.branch5x5_1.bn.weight), "branch5x5_1_bn_weights")
-    print_cpp_vector(torch.flatten(inception_a.branch5x5_1.bn.bias), "branch5x5_1_bn_bias")
-
-    print_cpp_vector(torch.flatten(inception_a.branch5x5_2.conv.weight), "branch5x5_2_conv_weights")
-    print_cpp_vector(torch.flatten(inception_a.branch5x5_2.bn.weight), "branch5x5_2_bn_weights")
-    print_cpp_vector(torch.flatten(inception_a.branch5x5_2.bn.bias), "branch5x5_2_bn_bias")
-
-    # branch3x3dbl
-    print_cpp_vector(torch.flatten(inception_a.branch3x3dbl_1.conv.weight), "branch3x3dbl_1_conv_weights")
-    print_cpp_vector(torch.flatten(inception_a.branch3x3dbl_1.bn.weight), "branch3x3dbl_1_bn_weights")
-    print_cpp_vector(torch.flatten(inception_a.branch3x3dbl_1.bn.bias), "branch3x3dbl_1_bn_bias")
-
-    print_cpp_vector(torch.flatten(inception_a.branch3x3dbl_2.conv.weight), "branch3x3dbl_2_conv_weights")
-    print_cpp_vector(torch.flatten(inception_a.branch3x3dbl_2.bn.weight), "branch3x3dbl_2_bn_weights")
-    print_cpp_vector(torch.flatten(inception_a.branch3x3dbl_2.bn.bias), "branch3x3dbl_2_bn_bias")
-
-    print_cpp_vector(torch.flatten(inception_a.branch3x3dbl_3.conv.weight), "branch3x3dbl_3_conv_weights")
-    print_cpp_vector(torch.flatten(inception_a.branch3x3dbl_3.bn.weight), "branch3x3dbl_3_bn_weights")
-    print_cpp_vector(torch.flatten(inception_a.branch3x3dbl_3.bn.bias), "branch3x3dbl_3_bn_bias")
-
-    # branchPool
-    print_cpp_vector(torch.flatten(inception_a.branch_pool.conv.weight), "branchPool_2_conv_weights")
-    print_cpp_vector(torch.flatten(inception_a.branch_pool.bn.weight), "branchPool_2_bn_weights")
-    print_cpp_vector(torch.flatten(inception_a.branch_pool.bn.bias), "branchPool_2_bn_bias")
-
-    input_shape = (1, 3, 8, 8)
-    input = torch.randn(input_shape)
-    print_cpp_vector(torch.flatten(input), "input")
-
-    output = inception_a(input)
-    output = torch.flatten(output)
-    print_cpp_vector(output)
-

examples/inception_v3/tests/inception_b.py

@@ -1,47 +0,0 @@
-import sys
-
-import torch
-from torchvision.models.inception import InceptionB
-
-sys.path.append("../../../tools")
-from utils import print_cpp_vector
-
-torch.manual_seed(0)
-
-@torch.no_grad()
-def init_weights(m):
-    if isinstance(m, torch.nn.Conv2d):
-        torch.nn.init.uniform_(m.weight)
-    elif isinstance(m, torch.nn.BatchNorm2d):
-        torch.nn.init.uniform_(m.weight)
-        torch.nn.init.uniform_(m.bias)
-
-with torch.no_grad():
-    inception_b = InceptionB(3)
-    inception_b.apply(init_weights)
-
-    # branch3x3
-    print_cpp_vector(torch.flatten(inception_b.branch3x3.conv.weight), "branch3x3_conv_weights")
-    print_cpp_vector(torch.flatten(inception_b.branch3x3.bn.weight), "branch3x3_bn_weights")
-    print_cpp_vector(torch.flatten(inception_b.branch3x3.bn.bias), "branch3x3_bn_bias")
-
-    # branch3x3dbl
-    print_cpp_vector(torch.flatten(inception_b.branch3x3dbl_1.conv.weight), "branch3x3dbl_1_conv_weights")
-    print_cpp_vector(torch.flatten(inception_b.branch3x3dbl_1.bn.weight), "branch3x3dbl_1_bn_weights")
-    print_cpp_vector(torch.flatten(inception_b.branch3x3dbl_1.bn.bias), "branch3x3dbl_1_bn_bias")
-
-    print_cpp_vector(torch.flatten(inception_b.branch3x3dbl_2.conv.weight), "branch3x3dbl_2_conv_weights")
-    print_cpp_vector(torch.flatten(inception_b.branch3x3dbl_2.bn.weight), "branch3x3dbl_2_bn_weights")
-    print_cpp_vector(torch.flatten(inception_b.branch3x3dbl_2.bn.bias), "branch3x3dbl_2_bn_bias")
-
-    print_cpp_vector(torch.flatten(inception_b.branch3x3dbl_3.conv.weight), "branch3x3dbl_3_conv_weights")
-    print_cpp_vector(torch.flatten(inception_b.branch3x3dbl_3.bn.weight), "branch3x3dbl_3_bn_weights")
-    print_cpp_vector(torch.flatten(inception_b.branch3x3dbl_3.bn.bias), "branch3x3dbl_3_bn_bias")
-
-    input_shape = (1, 3, 8, 8)
-    input = torch.randn(input_shape)
-    print_cpp_vector(torch.flatten(input), "input")
-
-    output = inception_b(input)
-    output = torch.flatten(output)
-    print_cpp_vector(output, "expected")

examples/inception_v3/tests/inception_blocks.py

@@ -0,0 +1,83 @@
+import sys
+
+import torch
+from torchvision.models.inception import (
+    InceptionA,
+    InceptionB,
+    InceptionC,
+    InceptionD,
+    InceptionE
+)
+
+sys.path.append("../../../tools")
+from utils import print_cpp_vector, export_model_weights
+
+torch.manual_seed(0)
+
+output_size = 50
+
+class InceptionBlockModel(torch.nn.Module):
+    def __init__(self, inception_block: torch.nn.Module, linear_in: int, *args, **kwargs) -> None:
+        super().__init__(*args, **kwargs)
+        self.inception_block = inception_block
+        self.fc = torch.nn.Linear(linear_in, output_size)
+
+    def forward(self, x):
+        x = self.inception_block(x)
+        x = torch.flatten(x)
+        x = self.fc(x)
+        return x
+
+
+@torch.no_grad()
+def init_weights(m: torch.nn.Module):
+    if isinstance(m, torch.nn.Conv2d):
+        torch.nn.init.uniform_(m.weight, -1, 1)
+    elif isinstance(m, torch.nn.BatchNorm2d) or isinstance(m, torch.nn.Linear):
+        torch.nn.init.uniform_(m.weight, -1)
+        torch.nn.init.uniform_(m.bias, 1)
+
+
+@torch.no_grad()
+def generate_module_test_data(m: torch.nn.Module, name: str):
+
+    print(name)
+    
+    input_shape = (1, 3, 4, 4)
+    input = torch.randn(input_shape)
+    print_cpp_vector(torch.flatten(input), "input")
+
+    m.eval()
+    inception_out = m(input)
+    linear_in = torch.flatten(inception_out).size(0)
+
+    inception_block = InceptionBlockModel(m, linear_in)
+    inception_block.apply(init_weights)
+
+    export_model_weights(inception_block, f"resources/{name}.bin")
+
+    inception_block.eval()
+    output = inception_block(input)
+    print_cpp_vector(torch.flatten(output), "expected")
+
+    print()
+
+
+if __name__ == "__main__":
+    m = InceptionA(3, 6)
+    generate_module_test_data(m, "inception_a")
+
+    m = InceptionB(3)
+    generate_module_test_data(m, "inception_b")
+
+    m = InceptionC(3, 64)
+    generate_module_test_data(m, "inception_c")
+
+    m = InceptionD(3)
+    generate_module_test_data(m, "inception_d")
+
+    m = InceptionE(3)
+    generate_module_test_data(m, "inception_e")
+
+
+

examples/inception_v3/tests/inception_c.py

@@ -1,73 +0,0 @@
-import sys
-
-import torch
-from torchvision.models.inception import InceptionC
-
-sys.path.append("../../../tools")
-from utils import print_cpp_vector
-
-torch.manual_seed(0)
-
-@torch.no_grad()
-def init_weights(m):
-    if isinstance(m, torch.nn.Conv2d):
-        torch.nn.init.uniform_(m.weight)
-    elif isinstance(m, torch.nn.BatchNorm2d):
-        torch.nn.init.uniform_(m.weight)
-        torch.nn.init.uniform_(m.bias)
-
-with torch.no_grad():
-    inception_c = InceptionC(3, 64)
-    inception_c.apply(init_weights)
-
-    # branch1x1
-    print_cpp_vector(torch.flatten(inception_c.branch1x1.conv.weight), "branch1x1_conv_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch1x1.bn.weight), "branch1x1_bn_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch1x1.bn.bias), "branch1x1_bn_bias")
-
-    # branch7x7
-    print_cpp_vector(torch.flatten(inception_c.branch7x7_1.conv.weight), "branch7x7_1_conv_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch7x7_1.bn.weight), "branch7x7_1_bn_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch7x7_1.bn.bias), "branch7x7_1_bn_bias")
-
-    print_cpp_vector(torch.flatten(inception_c.branch7x7_2.conv.weight), "branch7x7_2_conv_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch7x7_2.bn.weight), "branch7x7_2_bn_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch7x7_2.bn.bias), "branch7x7_2_bn_bias")
-
-    print_cpp_vector(torch.flatten(inception_c.branch7x7_3.conv.weight), "branch7x7_3_conv_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch7x7_3.bn.weight), "branch7x7_3_bn_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch7x7_3.bn.bias), "branch7x7_3_bn_bias")
-
-    # branch7x7dbl
-    print_cpp_vector(torch.flatten(inception_c.branch7x7dbl_1.conv.weight), "branch7x7dbl_1_conv_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch7x7dbl_1.bn.weight), "branch7x7dbl_1_bn_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch7x7dbl_1.bn.bias), "branch7x7dbl_1_bn_bias")
-
-    print_cpp_vector(torch.flatten(inception_c.branch7x7dbl_2.conv.weight), "branch7x7dbl_2_conv_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch7x7dbl_2.bn.weight), "branch7x7dbl_2_bn_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch7x7dbl_2.bn.bias), "branch7x7dbl_2_bn_bias")
-
-    print_cpp_vector(torch.flatten(inception_c.branch7x7dbl_3.conv.weight), "branch7x7dbl_3_conv_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch7x7dbl_3.bn.weight), "branch7x7dbl_3_bn_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch7x7dbl_3.bn.bias), "branch7x7dbl_3_bn_bias")
-
-    print_cpp_vector(torch.flatten(inception_c.branch7x7dbl_4.conv.weight), "branch7x7dbl_4_conv_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch7x7dbl_4.bn.weight), "branch7x7dbl_4_bn_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch7x7dbl_4.bn.bias), "branch7x7dbl_4_bn_bias")
-
-    print_cpp_vector(torch.flatten(inception_c.branch7x7dbl_5.conv.weight), "branch7x7dbl_5_conv_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch7x7dbl_5.bn.weight), "branch7x7dbl_5_bn_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch7x7dbl_5.bn.bias), "branch7x7dbl_5_bn_bias")
-
-    # branch_pool
-    print_cpp_vector(torch.flatten(inception_c.branch_pool.conv.weight), "branchPool_2_conv_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch_pool.bn.weight), "branchPool_2_bn_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch_pool.bn.bias), "branchPool_2_bn_bias")
-
-    input_shape = (1, 3, 8, 8)
-    input = torch.randn(input_shape)
-    print_cpp_vector(torch.flatten(input), "input")
-
-    output = inception_c(input)
-    output = torch.flatten(output)
-    print_cpp_vector(output, "expected")

examples/inception_v3/tests/inception_d.py

@@ -1,55 +0,0 @@
-import sys
-
-import torch
-from torchvision.models.inception import InceptionD
-
-sys.path.append("../../../tools")
-from utils import print_cpp_vector
-
-torch.manual_seed(0)
-
-@torch.no_grad()
-def init_weights(m):
-    if isinstance(m, torch.nn.Conv2d):
-        torch.nn.init.uniform_(m.weight)
-    elif isinstance(m, torch.nn.BatchNorm2d):
-        torch.nn.init.uniform_(m.weight)
-        torch.nn.init.uniform_(m.bias)
-
-with torch.no_grad():
-    inception_c = InceptionD(3)
-    inception_c.apply(init_weights)
-
-    # branch3x3
-    print_cpp_vector(torch.flatten(inception_c.branch3x3_1.conv.weight), "branch3x3_1_conv_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch3x3_1.bn.weight), "branch3x3_1_bn_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch3x3_1.bn.bias), "branch3x3_1_bn_bias")
-
-    print_cpp_vector(torch.flatten(inception_c.branch3x3_2.conv.weight), "branch3x3_2_conv_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch3x3_2.bn.weight), "branch3x3_2_bn_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch3x3_2.bn.bias), "branch3x3_2_bn_bias")
-
-    # branch7x7x3
-    print_cpp_vector(torch.flatten(inception_c.branch7x7x3_1.conv.weight), "branch7x7x3_1_conv_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch7x7x3_1.bn.weight), "branch7x7x3_1_bn_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch7x7x3_1.bn.bias), "branch7x7x3_1_bn_bias")
-
-    print_cpp_vector(torch.flatten(inception_c.branch7x7x3_2.conv.weight), "branch7x7x3_2_conv_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch7x7x3_2.bn.weight), "branch7x7x3_2_bn_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch7x7x3_2.bn.bias), "branch7x7x3_2_bn_bias")
-
-    print_cpp_vector(torch.flatten(inception_c.branch7x7x3_3.conv.weight), "branch7x7x3_3_conv_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch7x7x3_3.bn.weight), "branch7x7x3_3_bn_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch7x7x3_3.bn.bias), "branch7x7x3_3_bn_bias")
-
-    print_cpp_vector(torch.flatten(inception_c.branch7x7x3_4.conv.weight), "branch7x7x3_4_conv_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch7x7x3_4.bn.weight), "branch7x7x3_4_bn_weights")
-    print_cpp_vector(torch.flatten(inception_c.branch7x7x3_4.bn.bias), "branch7x7x3_4_bn_bias")
-
-    input_shape = (1, 3, 8, 8)
-    input = torch.randn(input_shape)
-    print_cpp_vector(torch.flatten(input), "input")
-
-    output = inception_c(input)
-    output = torch.flatten(output)
-    print_cpp_vector(output, "expected")

examples/inception_v3/tests/test_basic_conv2d.cpp

@@ -12,7 +12,7 @@ class BasicConv2dTest : public ::testing::Test {
     shape2d     kernelSize;
     shape2d     stride;
     shape2d     padding;
-    std::string prefix = "test";
+    std::string prefix = "test.";
 
     float *d_input;
     float *d_output;
@@ -46,7 +46,7 @@ class BasicConv2dTest : public ::testing::Test {
         std::pair<std::string, CUDANet::Layers::SequentialLayer *> layerPair =
             basic_conv2d->getLayers()[0];
 
-        ASSERT_EQ(layerPair.first, prefix + ".conv");
+        ASSERT_EQ(layerPair.first, prefix + "conv");
 
         CUDANet::Layers::Conv2d *conv =
             dynamic_cast<CUDANet::Layers::Conv2d *>(layerPair.second);
@@ -60,7 +60,7 @@ class BasicConv2dTest : public ::testing::Test {
         EXPECT_EQ(cudaStatus, cudaSuccess);
 
         layerPair = basic_conv2d->getLayers()[1];
-        ASSERT_EQ(layerPair.first, prefix + ".bn");
+        ASSERT_EQ(layerPair.first, prefix + "bn");
 
         CUDANet::Layers::BatchNorm2d *bn =
             dynamic_cast<CUDANet::Layers::BatchNorm2d *>(layerPair.second);

examples/inception_v3/tests/test_fixture.hpp

@@ -0,0 +1,50 @@
+#ifndef TEST_FIXTURE_H
+#define TEST_FIXTURE_H
+
+#include <cmath>
+#include <gtest/gtest.h>
+
+#include "inception_v3.hpp"
+
+class InceptionBlockTest : public ::testing::Test {
+  protected:
+    CUDANet::Model *model;
+
+    cudaError_t cudaStatus;
+
+    shape2d inputShape;
+    int     inputChannels;
+
+    int outputSize;
+
+    std::vector<float> input;
+    std::vector<float> expected;
+
+    virtual void SetUp() override {
+        model = nullptr;
+    }
+
+    virtual void TearDown() override {
+        // Clean up
+        delete model;
+    }
+
+    void runTest() {
+        EXPECT_EQ(
+            input.size(), inputShape.first * inputShape.second * inputChannels
+        );
+
+        float *output = model->predict(input.data());
+
+        cudaStatus = cudaGetLastError();
+        EXPECT_EQ(cudaStatus, cudaSuccess);
+
+        EXPECT_EQ(outputSize, expected.size());
+
+        for (int i = 0; i < outputSize; ++i) {
+            EXPECT_NEAR(expected[i], output[i], 1e-3f);
+        }
+    }
+};
+
+#endif

examples/inception_v3/tests/test_inception_e.cpp

@@ -0,0 +1,69 @@
+#include <gtest/gtest.h>
+
+#include <inception_v3.hpp>
+
+#include "test_fixture.hpp"
+
+class InceptionEModel : public CUDANet::Model {
+  public:
+    InceptionEModel(
+        const shape2d inputShape,
+        const int     inputChannels,
+        const int     outputSize
+    )
+        : CUDANet::Model(inputShape, inputChannels, outputSize) {
+        inception_e =
+            new InceptionE(inputShape, inputChannels, "inception_block.");
+        addLayer("", inception_e);
+        fc = new CUDANet::Layers::Dense(
+            inception_e->getOutputSize(), 50,
+            CUDANet::Layers::ActivationType::NONE
+        );
+        addLayer("fc", fc);
+    };
+
+    float *predict(const float *input) override {
+        float *d_input = inputLayer->forward(input);
+        d_input        = inception_e->forward(d_input);
+        d_input        = fc->forward(d_input);
+        return outputLayer->forward(d_input);
+    }
+
+  private:
+    InceptionE             *inception_e;
+    CUDANet::Layers::Dense *fc;
+};
+
+TEST_F(InceptionBlockTest, InceptionETest) {
+    inputShape    = {4, 4};
+    inputChannels = 3;
+    outputSize    = 50;
+
+    model = new InceptionEModel(inputShape, inputChannels, outputSize);
+    model->loadWeights("../tests/resources/inception_e.bin");
+
+    input = {1.85083f,  0.11234f,  0.05994f,  -1.02453f, 0.21205f,  -0.67387f,
+             0.66981f,  -0.40378f, 0.34194f,  0.92048f,  0.87556f,  0.81094f,
+             -1.55728f, -0.70326f, -0.25078f, -0.10276f, 1.10463f,  -2.40992f,
+             -1.7226f,  -0.18546f, 0.14397f,  -1.24784f, -0.35248f, -1.28729f,
+             0.44803f,  1.68539f,  -1.05037f, 0.32115f,  -0.12896f, 1.02391f,
+             0.95329f,  -0.81876f, -1.05828f, 0.09282f,  -0.38344f, 2.05074f,
+             2.1034f,   1.65832f,  1.63788f,  -1.32596f, -1.43412f, -1.28353f,
+             0.70226f,  0.9459f,   0.8579f,   0.15361f,  0.34449f,  -1.70587f};
+
+    expected = {1614.15283f,   -11319.01855f, 614.40479f,    5280.0293f,
+                1914.45007f,   -2937.50317f,  -11177.16113f, 3215.01245f,
+                6249.16992f,   5654.91357f,   -11702.27148f, 13057.32422f,
+                8665.35742f,   3911.11743f,   5239.45947f,   -11552.88477f,
+                -8056.7666f,   -16426.19922f, -1383.04346f,  6573.53125f,
+                -12226.16992f, -6641.0957f,   -9614.80078f,  -9313.30273f,
+                7023.68848f,   2089.5752f,    1095.53369f,   -1387.65698f,
+                -7928.21729f,  -9489.18848f,  4159.78613f,   -690.03442f,
+                -8356.81738f,  12364.08203f,  8226.95703f,   8822.66602f,
+                -5462.90381f,  -1037.42773f,  12958.68555f,  -666.58423f,
+                2032.38574f,   -9534.14062f,  -947.41333f,   689.37158f,
+                4585.76465f,   -23245.36719f, 975.83398f,    -1253.45703f,
+                -14745.35059f, -2588.05493f};
+
+    runTest();
+}